What we're reading this week

Reframes the shrinking attention span as a structural class divide rather than a personal failing, grounded in industry data (shrinking theatrical releases, acts trimmed 'to get to the point').

Original longitudinal research (the source of the 'velocity up, DevEx down' finding) showing productivity and developer experience decoupling under AI, via MacIntyre's internal vs. external goods.

Mollick's reframe of the agentic shift: the human role becomes managing rather than chatting, with warts-and-all examples (a permissions gotcha where ChatGPT auto-sent an email).

A documented agentic Claude build whose core move inverts the AI from ghostwriter to interviewer, so the human articulates rather than outsources their thinking.

A practitioner who inherits abandoned big-company codebases argues against Naur's theory-building absolutism to defend partial understanding, where 'knowing enough to guide' diverges from full comprehension.

Why coding-agent benchmarks grade 'did the test pass' and not 'is the code good'. Models improve at one-off tickets but not maintenance, because that is the only thing benchmarks can score.

First-principles reasoning about hardware limits to distrust benchmarks and out-build a market; turbopuffer hit $100M ARR on under $1M raised and cut Cursor's search bill 95%.

Netflix's CPTO argues systems thinking, not specialist execution, is the durable skill as AI collapses execution cost, and treats AI fluency as a universal expectation.

Sam Rose's warts-and-all account of a near-fully LLM-authored 100K-line project made trustworthy only through line-by-line review and 204 differential tests against real k3s. Verification and taste are the scarce inputs.

Companion walkthrough detailing the voice-guide file, Writer's Council scoring personas, and reinforcement loop; 'yap to text' forces you to articulate specifics.

Ben Thompson's first-principles economic teardown (COGS vs R&D, commodity marginal-cost dynamics) of where value concentrates as model-creation costs collapse. The strongest piece of the batch, and it led both the Matter and Hacker Newsletter digests.

Kevin Kelly's canonical framework: when copying costs collapse to zero, eight 'generatives' (trust, immediacy, authenticity, findability…) map exactly what stays scarce. Timeless, and it reads as prophetic in the AI era.

Antirez (Redis creator) argues that when AI writes the code, human leverage shifts to owning the ideas and design, with a DESIGN.md capturing each data structure's ideas so contributors own them first.

Stripe's Kevin Yien on what's scarce once every PM can build: judgment, forced opinion, and 'deep generalist' lived experience, plus encoding that judgment into a shared 'team brain.'

Data-grounded read of DX's Q2 2026 report: AI productivity gains concentrate among small orgs into a steeper power law, while DevEx declines and AI spend (up 28x) outruns flat innovation.

The bottleneck shifts from writing code to reviewing it as AI output surges, with the sharper worry that review-with-intent is eroding into rubber-stamping AI-approved PRs.

A practitioner who ran 'lights-off' agentic delivery grounds a structural argument (benchmarks reward passing tests but have no oracle for maintainability, so judgment on design stays scarce) in a concrete 4-phase workflow.

Ronacher on the unifying question: as AI collapses execution cost, people hand machines the very judgment, attention, and meaning they should hold onto, feeding a race with no finish line.

Names the exact gap between adequate and genuinely good work, drawn from coaching retrospectives, and hands you a reusable diagnostic ('This is defensible. But is it good?').

Beej's self-aware reflection on 'making vs. asking-to-be-made', using disclaimed AI art, carpentry, and code to probe what stays irreducibly yours when execution is delegated.

Reframes career desire from optimizing outcomes toward becoming a person you'd be proud of, with one operationalizable move: define values as 'in situation X I'd be proud to respond by Y.'

Practitioner-researched excerpts (DJ DiDonna) on sabbaticals as transformation: the Achiever/Seeker/Explorer archetypes, and a critique of turning rest into another accomplishment checklist.

A solo maker's clear-eyed retrospective on ending a hand-built daily puzzle: the economics of the unscalable, and an honest framework for when a passion project's passion has corroded enough to stop.

Will Larson's structural feedback-loop insight: a team's growing effectiveness converts invisible 'suppressed demand' into real demand, so fixing a team can feel like nothing improved.

Leadership composure as a 'broadcast' that's a separate job from solving the problem, stress-tested with a cross-domain pairing (a calm cockpit announcement vs. denial aimed at people who must act).

Frames project management as structural laws rather than opinion. A backlog encodes the ignorance present when it was written; the board is the single artifact that gives everyone the same project.

Framing the problem space (boundaries, metrics, constraints) is the work, with a transferable method: question dump, then triage the questions by 'power.'

A deliberately 'human postmortem' of the 1996 AOL outage, exposing what the formal five-whys/SLO format designs out of view: the unequal human impact.

A cross-domain conversation on what stays with humans as execution gets cheap: meta-intelligence over raw smarts, and why removing the human hides accountability rather than removing it.

Maps a biological mechanism (a cavefish suppressing sight when the environment stops rewarding it) onto why companies stop recognizing competence when incentives no longer express the trait.